Why Are the Different Presidential Forecasts So Far Apart?
|
Version 0 of 1. Hillary Clinton currently has a 71 percent chance of winning the presidency, according to The Upshot’s forecasting model. This is down from 90 percent last month, but higher than some other models we’re tracking, which put the odds between 58 percent and 85 percent. Part of the discrepancy comes from the use of different information. The PredictWise number — 74 percent — incorporates a sharp jump in betting markets that occurred during the first presidential debate. This jump, if it’s real, is not yet reflected in polls, which take days to conduct. But most of the difference has to do with different model assumptions. Poll-based models need to take a position on two key questions: How useful are older polls? And to what extent do states move together? The answer to the first question informs how heavily the model weights new information. Weighting new polls very heavily means you can be quicker to respond to new trends and to pick up turning points, but it also makes your forecast more unstable and more likely to react to noise (what we like to call “chasing shiny objects”). Ideally a forecast finds a balance between these two extremes, trading some quickness for stability. The second question informs how much a shift in one state causes shifts in the others. If all the states were perfectly correlated and moved uniformly, a two-point swing toward Mrs. Clinton in South Carolina would mean a two-point swing toward Mrs. Clinton in Maine, Kansas and everywhere else. If the states are totally uncorrelated, movements in South Carolina would tell you nothing at all about Maine. On both of these issues — correlation and shiny-object-chasing — our model chooses a middle value, based on how presidential polling has moved in past elections since 1980. But our choices aren’t the only ones possible. Different choices for these parameters lead to different forecasts, and different views of how the race has shifted. The chart above shows the number of points we would have said Mrs. Clinton was leading by in Florida, based on different — but not unreasonable – assumptions about these questions. Generally speaking, the choices don’t matter that much: There is about a two-point difference in some weeks. And a two-point spread across different parameter choices is nothing to be afraid of, especially amid other types of error. Remember: The whole point of a model is to allow people to step back from daily overreactions to the latest polling, and to give a more complete picture of where the race stands. But two points is not nothing, either. If each state shifted by two points toward Mr. Trump, for example, it would transform Mrs. Clinton’s slight lead into a dead heat. The lineup of states right now gives Mr. Trump a slight Electoral College advantage: It’s easier for him to win the Electoral College while losing the popular vote than it would be for Mrs. Clinton, so he doesn’t have to pull even in the popular vote estimate in order to have an even chance of winning the election. As my colleague Nate Cohn wrote the other day, uncertainty in the race remains high. There are 13 states that we classify as “leaning” or “tossups,” meaning there is at least a 15 percent chance that each will be won by the underdog. That’s a historically high number of states to be that uncertain about this far into the campaign, and the various forecasting models have interpreted the previous month somewhat differently. As the post-debate polls come in over the next few days, I expect the models to converge toward one another once more. In general, the more certainty there is, the less that the choices of the individual models matter. But even if Mrs. Clinton gains in the polls over the next few weeks, the effects immediately after a debate are often short-lived. Our model will take the post-debate polling at face value, but there is perhaps an argument to be made that applying a post-debate correction to the polling numbers is worthwhile. For a period right after each convention, we used such a correction, essentially giving those polls a touch less weight. The result was that our model showed less bounce than it otherwise would have — taking extra steps to resist the shiny object. |